The AI assessment: What it delivers, and what it costs when you skip it

An AI assessment is a structured examination of where AI can take over work in an organisation, and it should end in three things:

• Scored list of opportunities,
• Shortlist of three to five use cases with a business case,
• Roadmap with owners and budget.

If all you receive is a report with recommendations, you bought an inventory rather than an assessment. That distinction sounds semantic, and it is where most of the money disappears.

At DataNorth AI we have run more than forty AI Assessments over the last 18 months, across municipalities, manufacturers, construction firms, healthcare organisations and investment houses. There is enough repetition in that to name a pattern, and it is not the one most boards expect: we almost never meet an organisation short of ideas about AI, but constantly meet organisations with far too many of them and no mechanism for saying no.

What follows is what an assessment involves, what a good one produces, how to recognise a weak one and what it should cost. That should be enough to judge whether the proposal on your desk is worth signing, even when it did not come from us.

What is an AI assessment, and what is it not?

An AI Assessment starts by examining processes rather than technology. An assessment starts by identifying and analyzing where people’s time actually goes and which part of that can be structured, and it ends with a defensible decision about where to put money and attention first. Technology comes afterwards, because which model or platform fits can only be settled once it is clear what needs building.

Three things get confused with it regularly. An AI strategy looks at your market and positioning and answers what AI does to your business model. An AI workshop brings knowledge and momentum into a team and is excellent at surfacing ideas, but it produces no shortlist and no business case. An AI Proof of Concept shows that something is technically possible, which is a different question from whether it is worth doing at your organisation.

Buying an assessment when you actually need one of those three means paying for the wrong product. That happens more often than vendors admit, and the section on when to skip it deals with that directly.

What should a good AI assessment deliver?

Four documents you can use without translation, and they build on each other in that order.

It starts with a scored longlist: every opportunity surfaced in the interviews, each carrying a value score and a complexity score. How many that comes to depends on size. In our engagements it is typically 70 to 140, with 130 at a clothing brand across nine departments and 74 at a construction firm with eight. A list like that is not a deliverable in itself, but it is the evidence that nothing was missed, and you need that evidence the moment somebody asks six months later why their process was left out.

From the longlist comes a shortlist of three to five use cases, each with a stated reason why the rest fell away. That shortlist should be visible across the organisation, because a use case that only touches the department that proposed it rarely generates enough momentum to justify a second project. So ask not only what was selected, but who notices when it works.

Each selected use case then gets a business case: the hours currently going into the process, the expected saving or revenue, the estimated build cost and the payback period. Those numbers can carry a wide range, because an estimate with margins survives a board meeting and a feeling does not. What you are buying is not precision, it is the ability to weigh one use case against another.

Everything lands in a twelve-month roadmap, with one owner per use case, a budget, and a point at which you measure whether it worked. That last part is not a formality. Without an owner who has time allocated, a roadmap is a wish list with dates attached, and it does not survive the first busy month.

Why do AI projects stall without one?

Because they start with the technology rather than the process, and because nobody defined at the outset what “it works” would mean. In the engagements we take over from an earlier attempt, those two causes almost always appear together. A tool was chosen, there was genuine enthusiasm, something got built that demonstrably functions, and yet it never reaches production. Ask which number was supposed to move and the room goes quiet.

That observation is not ours alone. Gartner predicts that over 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The interesting part is what the list leaves out. None of those three causes concerns whether the model is good enough. All three are scoping problems, and scoping is exactly the work an assessment does.

The practical consequence is less obvious than it looks. If the cause sits in scoping, then inventorying opportunities is not the fix, because it widens the field of choice without making the choice any easier. A longlist of 130 opportunities can paralyse an organisation as effectively as having no list at all. What makes the difference is not the length of the list but the gate behind it.

What should the process look like?

Five steps, with the weight in the interviews rather than in writing the report. Across five departments that should fit into two to four weeks.

1. Inventory. Roughly an hour per department, with the people who do the work rather than the leadership team. The question can be blunt: where does most of your time go on work you dislike?
2. Scoring. Every opportunity gets a value score and a complexity score, for instance on a scale of 1 to 3. High value and low complexity move forward, the rest waits with a reason attached.
3. The feasibility gate. A fixed set of conditions a use case must meet before anything gets built. A no here should mean parked rather than killed, because circumstances change.
4. Business case. Hours, cost, benefit and payback per surviving use case, with margins wherever the estimate is uncertain.
5. Roadmap. What starts when, who owns it, what it costs and what you will measure it against.

Where it goes wrong in practice: the feasibility gate

Step three is the one most often skipped and the one worth pushing hardest on. Without a gate you pick the use case with the most energy behind it, and that is rarely the one that is ready to build. Enthusiasm and readiness are different things, and they often live in different departments.

A workable gate is short and strict. In our own engagements we hang seven conditions under it, and they are perfectly reproducible without us:

1. The process is documented as it runs today, and as it should run afterwards.
2. A business owner is named, with time actually allocated.
3. The data exists and is accessible.
4. Integrations and access have been tested, with a date attached.
5. There is a success measure with a baseline and a target.
6. There is a test set you can measure against.
7. Scope is agreed, including the AI Act risk class.

A gate is only worth something once it is attached to something. At one of our clients it operates inside a Kanban flow running from backlog through nominated, in preparation and ready, into development, pilot, production and embedded, where a use case that fails the gate moves backwards rather than forwards. There is a less obvious rule alongside it: keep at least two approved use cases in stock per developer. Without that buffer, somebody eventually starts on a half-prepared use case because nothing else was queued, and at that point the gate has been opened in practice regardless of what the board says.

This is ultimately the part you are paying an assessment for. Anyone can produce the list, and a good workshop gets you a long way towards it internally. The discipline to say no to 125 of 130 opportunities, and to hold that no six months later, is what an assessment adds.

How do you spot a weak AI assessment?

By the deliverable, not by the process or the name on the cover. Four signals hold up almost every time:

• Nothing gets cut. If every opportunity is labelled promising, the work was an inventory rather than an assessment, and the choice still sits entirely with you.
• There are no numbers. “Significant time savings” does not survive a board meeting. Without hours and euros there is nothing to test an investment against.
• The platform is settled in chapter one. That makes the assessment a quote with a preface, and the conclusion was never in doubt.
• The assessor is also the builder, unstated. That combination can work very well and we do it ourselves, but it belongs on the table explicitly, alongside an honest view of which use cases your own team or an existing vendor should handle instead.

So when a proposal arrives, ask about the deliverables first and then about how many opportunities are expected to fall away. The answer to that second question tells you more about quality than the page count of the report.

What does skipping an AI assessment actually cost?

Four bills, and you only see them once they are already running.

The most expensive is the pilot nobody owns. A proof of concept that works technically but where nobody can define what “good” means does not reach production, and the loss runs well past the budget line. The team that spent months on it loses its appetite, and the next attempt starts with less internal backing than the first. Assume six to twelve months of lost time per failed attempt, and an organisation that is measurably harder to move afterwards.

Running alongside that is the shadow AI bill, which nobody opens until something goes wrong. With no inventory and no policy, people choose their own tools, and it happens faster than most boards assume: in nearly every assessment we run, at least one department turns out to be paying for a subscription IT does not know about. That is not obstruction, it is a symptom of the official route being slower than the work. The financial exposure is now well documented, with IBM’s 2025 Cost of a Data Breach research finding that a high level of shadow AI added roughly 670,000 US dollars to the global average breach cost.

The same absence of oversight produces the third bill, tool sprawl. Each department buys its own licence, you pay three times over for overlapping functionality, and nobody owns the data flows between them. Individually these are small numbers. Collectively they recur every month and nobody cancels them, because nobody can say with confidence who is paying for what.

The fourth bill is the newest. Article 4 of the EU AI Act requires any organisation deploying AI systems to take measures supporting AI literacy among its staff. That duty has applied since 2 February 2025, and enforcement by national authorities began on 2 August 2026. The law prescribes no particular course or certificate, but it does expect you to evidence the effort, and you cannot do that without knowing which AI systems are in use. Which is why a risk class per use case belongs in an assessment as standard: it is the same inventory work, and you need it anyway.

What should an AI assessment cost?

Market pricing varies widely and says little about quality, so ask about the deliverables and the timeline first and about the price second. As a guideline, an assessment covering five departments over two weeks should land somewhere in the region of 15,000 euros, plus a few thousand per additional department. If a proposal comes in far below that, check whether the interviews are genuinely included. If it comes in far above, ask what the extra buys.

That is also what we charge: 15,000 euros for five departments, plus 2,500 euros per additional department. We publish the figure while many providers will not name one before the third conversation, and there is a practical reason for that. An assessment is rarely a standalone purchase, so you need to budget the route after it before you say yes at the start. For orientation, a one-hour executive demo costs 1,500 euros with us, a four-hour AI workshop for up to ten participants 2,900 euros, and a proof of concept on the first selected use case 6,000 to 24,000 euros depending on whether the build takes one week or four. At one of our customers that was a forty-hour proof of concept followed by ten weekly sprint demos, after which their AI assistant went live on WhatsApp Business.

Set that order of magnitude against one stalled pilot and the arithmetic becomes straightforward. An assessment costs you two weeks and roughly half the price of a failed pilot. A pilot that dies on a use case that should never have been on the list costs you six months and your internal credibility. That gap is the main reason this step exists.

When do you not need an AI assessment?

More often than vendors will tell you. There are three situations where we actively advise against it, and that is arithmetic rather than modesty: in these cases you would be paying for confirmation.

1. The first is that you can already name one clearly bounded problem, such as manual order intake or a knowledge base nobody searches. Go straight to a proof of concept on that single process. An assessment will mostly tell you what you already knew, at the cost of two weeks you could have spent building.
2. The second is that you have fewer than ten employees. The inventory fits in an afternoon and the formal steps cost more than they return. Start with a workshop and see what surfaces.
3. The third is that you are mid-migration on your ERP or core systems. Wait until the data flows settle, because an assessment of a landscape that changes in six months ages before the roadmap begins. The opportunities will still be there, and the feasibility gate in step three produces a very different answer once the migration is done.

Where the assessment sits in the whole

An AI assessment is an intermediate step rather than a beginning or an end. An AI demo or AI workshop often comes before it to get an organisation moving, and an AI proof of concept follows to actually build the first selected use case. You do not have to walk that route from the start, but you do need to know where you stand, because that determines whether an assessment moves you forward or simply confirms what you already knew.

If you are not sure, an hour with somebody who has run dozens of them costs less than an assessment you did not need. That holds with us, and it holds just as well at another firm.

Frequently asked questions (FAQ) about AI Assessments